At the moment, we create a regular grid of points spaced 2km apart, that is "crossed" with the training BGC units and a DEM (at the moment a NAmerica DEM at 250m).
Because some BGC units are much larger than others and RF models are sensitive to highly unbalanced data, we then balanced the data to reduce the number of training points in the largest BGC units.
We may want to revise this strategy and, instead apply sampling approach where we sample the same (or a similar) number of points for each unit at the start - this would result in smaller units having more spatially clumped samples than the larger units.
At the moment, we create a regular grid of points spaced 2km apart, that is "crossed" with the training BGC units and a DEM (at the moment a NAmerica DEM at 250m). Because some BGC units are much larger than others and RF models are sensitive to highly unbalanced data, we then balanced the data to reduce the number of training points in the largest BGC units.
We may want to revise this strategy and, instead apply sampling approach where we sample the same (or a similar) number of points for each unit at the start - this would result in smaller units having more spatially clumped samples than the larger units.